ikawrakow/ik_llama.cpp
1
0
Fork 0
mirror of https://github.com/ikawrakow/ik_llama.cpp.git synced 2026-02-22 22:24:11 +00:00
Code Issues Packages Projects Releases Wiki Activity

IQ4_K_R4 (#138)

Browse Source 
* iq4_k_r4: WIP

* iq4_k_r4: Zen4 and hopefully AVX2

On Zen4 we get PP-512(LLaMA-3.1-8B) = 232.6 t/s, up from 182.2 t/s
for iq4_k. Applying the extra shift costs a ~6 performance penalty.

* iq4_k_r4: AVX2

PP-512 = 227.60 t/s. The shifts are really costly.

* iq4_k_r4: NEON

We get PP-512(LLaMA-3.1-8B) = 108 t/s, up from 58.2 t/s for iq4_k.

---------

Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
This commit is contained in:
Kawrakow
2024-12-12 16:04:20 +01:00
committed by GitHub
parent 66ade83e56
commit ce97b0325e
10 changed files with 443 additions and 16 deletions
Download Patch File Download Diff File Expand all files Collapse all files

2
ggml/include/ggml.h
View File

@@ -421,6 +421,7 @@ extern "C" {
GGML_TYPE_IQ4_XS_R4 = 223,
GGML_TYPE_Q6_0_R4 = 233,
GGML_TYPE_IQ2_BN_R4 = 335,
GGML_TYPE_IQ4_K_R4 = 339,
GGML_TYPE_COUNT,
};
@@ -492,6 +493,7 @@ extern "C" {
GGML_FTYPE_MOSTLY_IQ4_XS_R4 = 222, // except 1d tensors
GGML_FTYPE_MOSTLY_Q6_0_R4 = 227, // except 1d tensors
GGML_FTYPE_MOSTLY_IQ2_BN_R4 = 329, // except 1d tensors
GGML_FTYPE_MOSTLY_IQ4_K_R4 = 332, // except 1d tensors
};
// available tensor operations: